Discursive-Network Model Framework

• The Discursive-Network Model is a framework that defines discourse as networked communications, where publications and textual traces yield emergent codification.
• It employs methodologies from network analysis, information theory, and latent semantic modeling to reveal structured relations and dynamic feedback in discursive systems.
• The model bridges scientific communication and social media by integrating observable ties with latent semantic organization to map complex interaction patterns.

Searching arXiv for relevant papers on discursive-network models and adjacent formulations. arxiv_search(query="discursive network model scientific communication discourse networks citations codification", max_results=10) The Discursive-Network Model is a conceptual, analytical, and methodological framework in which discourse is treated as a networked system of communications whose elementary operations are publications, utterances, documents, or other discursive traces, and whose higher-order organization emerges from patterned relations among them. In its classical scientific-communication formulation, the model holds that “discursive knowledge emerges as codification in flows of communication,” and that these flows are constrained and enabled by historically instantiated networks of texts, references, journals, words, and other textual attributes (0911.1308). Across later variants, the same general orientation recurs: discourse is not reduced either to individual cognition or to isolated texts, but is modeled as a structured field of relations in which latent semantic organization, observed ties, and recursive feedback jointly shape what can be said, recognized, stabilized, and circulated (Leydesdorff, 2012). In this sense, the Discursive-Network Model names not a single canonical formalism but a family of approaches that represent discourse as relationally organized, historically retained, and, in some variants, computationally operationalized through graphs, matrices, entropy measures, or coupled latent-state systems (0911.1308, Srivatsan et al., 2018, Hansen et al., 2020).

1. Historical and theoretical formulation

The most explicit general formulation is given by Lucio-Arias and Leydesdorff, who define scientific knowledge as codification in communication flows rather than as something primarily located in individual minds, laboratories, or institutions (0911.1308). In that account, scientific publications are the micro-operations of the science system. They are the events through which the system reproduces itself autopoietically. Communication flows comprise publications, submissions, validations, references, reinterpretations, and recombinations; networks of communications are the historically instantiated structures left by those flows at any given moment, including citation networks, clusters of journals, co-word constellations, and specialty structures (0911.1308).

A decisive distinction in this tradition is between first-order social relations among authors and institutions, and second-order relations among documents and textual attributes (0911.1308). Authors form social networks, but documents form an archive with self-referential dynamics. Texts cite previous texts, are cited in turn, and can be grouped by words, references, journals, or themes. Discursive knowledge is therefore defined as knowledge generated in a second-order dynamics among texts, and codification is the reflexive stabilization of meaning within these communication flows (0911.1308).

Leydesdorff later reformulated the same architecture as the interaction of two coupled topographies: network space, the space of observable relations among nodes, and vector space, the space of latent semantic dimensions inferred from distributions and correlations (Leydesdorff, 2012). In that formulation, social network analysis addresses information-processing in network space, whereas latent semantic analysis addresses meaning-processing in vector space (Leydesdorff, 2012). This distinction is foundational for the Discursive-Network Model because it prevents discourse from being collapsed into either mere co-occurrence or mere social tie structure.

A related but distinct philosophical extension appears in work on online misinformation, where discourse is reconceptualized as a socially structured practice of circulation, signaling, and exclusion rather than as a neutral carrier of truth-conditions. In that account, online information is “better understood as gossip,” and its circulation is linked to a “discursive grammar” organized by irresponsible speech, unwarranted assertion, exclusion of the target, and masked implication and import (Bourbon et al., 2022). This suggests that discursive-network approaches can be extended beyond scientific communication into socially embedded, identity-laden, and affectively structured environments.

2. Core ontology: documents, relations, codification, and specialties

In the classical model, a publication is simultaneously positioned in a network of formalized relations, located along a temporal trajectory of previous publications, and capable of retrospectively redefining the meaning of prior work from hindsight (0911.1308). References are not merely links from one paper to another; they are mechanisms by which a text selects and reconstructs previous discourse. Citation relations instantiate network structure at a given moment, while references in new texts recursively update the archive (0911.1308).

The model specifies three interacting selection mechanisms or selection environments (0911.1308). First, there is selection by structure at each moment of time, that is, the synchronic positioning of a publication in a network of citation or other textual relations. Second, there is historical shaping of the network into a system along a trajectory, that is, the diachronic retention and path-dependent stabilization of sequences of publications. Third, there is reflexive restructuring by further communications: new publications can reinterpret prior communications from hindsight, changing their significance and reconstructing discourse itself (0911.1308). This triadic structure is the core dynamic of the model.

Codification is tied to references, citations, textual exchanges, and jargon formation (0911.1308). Repeated selections among claims, references, and symbols produce more specialized and controlled repertoires, allowing scientific communities to process complexity beyond ordinary language. As codification deepens, differentiation increases and specialties emerge with increasingly specific codes and criteria (0911.1308). These specialties are not treated as simple hierarchical subdivisions; rather, as they gain autonomy, they develop in “orthogonal directions” (0911.1308). A plausible implication is that the model treats the organization of knowledge as multidimensional rather than merely taxonomic.

The relation between bottom-up emergence and top-down control is equally central. New claims originate bottom-up in local practices and enter discourse as publications, thereby providing variation. Once publications accumulate, latent intellectual structures emerge at the network level—specialties, paradigms, criteria of validity, codified repertoires, journal structures, and concept-symbol systems—which then feed back top-down on future publications (0911.1308). There is “no single center of control among the flows of communication,” but codification nonetheless shapes subsequent bottom-up production (0911.1308).

3. Relational and positional variants

A persistent problem in discursive-network analysis is how to relate observable ties to latent semantic structure. Leydesdorff’s distinction between network space and vector space gives one influential answer: observable networks correspond to co-authorship, co-word ties, citation links, or other direct relations, while latent dimensions correspond to discursive positions, themes, frames, or codes reconstructed from distributional patterns (Leydesdorff, 2012). Two entities can thus be close semantically without being directly linked, just as they can co-occur without occupying the same semantic position (Leydesdorff, 2012).

This distinction reappears in empirical work that starts from actor–concept or actor–topic incidence and then constructs projected networks. In the study of UN Security Council speeches on Afghanistan, discourse is operationalized first as latent thematic structure recovered by LDA and then as “speaker-topic relations,” where weighted ties connect speakers to topics according to aggregated topic assignments across speeches (Schoenfeld et al., 2018). The resulting two-mode network is thresholded and projected into a one-mode country-country network, where recognizable coalitions and geopolitical divisions emerge (Schoenfeld et al., 2018). The model therefore treats discourse as a relation between actors and latent themes, with coalition structure inferred from shared thematic orientation.

A related socio-semantic architecture appears in the analysis of Italian Covid-19 Twitter discourse, which combines an actor layer derived from retweet behavior and a concept layer derived from hashtag co-usage (Mattei et al., 2021). There, the semantic network is a validated projection of a bipartite user–hashtag network, and discursive communities are first inferred from verified–non-verified retweet patterns and then compared in terms of exposure to semantic communities and d/misinformation arguments (Mattei et al., 2021). This is very close to a multiplex actor–actor / actor–concept / concept–concept formulation.

Other work reconstructs discourse networks directly from shared participation in latent narratives. In cross-platform discourse modeling, users are represented by TF-IDF-weighted vectors over discovered narrative clusters, and user-user edges are induced by cosine similarity of those vectors (Gerard et al., 22 May 2025, Gerard et al., 10 Oct 2025). In such formulations, the network is not observed directly through follows, mentions, or reposts; it is inferred from shared narrative engagement (Gerard et al., 10 Oct 2025). This suggests a broader family resemblance: discursive-network models frequently pass through an implicit or explicit actor–discourse incidence layer before projecting to actor-actor structure.

4. Formalization and measurement

The original scientific-communication model is explicitly information-theoretical. Shannon entropy for a variable $x$ is given as

$H_x = - \sum_x p_x \log_2 p_x$

and for a joint distribution

$H_{xy} = - \sum_{x,y} p_{xy} \log_2 p_{xy}.$

Its central multivariate measure is the configurational information among three dimensions,

$$\mu^*_{xyz} = H_x + H_y + H_z - H_{xy} - H_{xz} - H_{yz} + H_{xyz}.$$

Negative configurational information is interpreted as a reduction of uncertainty at the level of the configuration, that is, system-level structuration or synergy (0911.1308).

The same paper places this formalization in a “triple helix” internal to scientific communication: cognitive/content, textual/journal, and social/scholarly-institutional dimensions (0911.1308). Theories, texts, and scholars or institutions interact recursively, and configurational information measures whether those interactions produce system-level synergy (0911.1308). A plausible implication is that discursive-network analysis, in this version, becomes an empirical program for detecting latent intellectual organization through multivariate uncertainty reduction.

A different formalization appears in opinion dynamics on discourse sheaves, where actors are vertices, communication channels are edges, private opinion spaces are attached to vertices, and public discourse spaces are attached to edges (Hansen et al., 2020). The core condition of local consistency on an edge $e=u\sim v$ is

$\Fc_{u\face e}x_u=\Fc_{v\face e}x_v,$

and the sheaf Laplacian is

$L_\Fc = \delta^T\delta.$

Here the Discursive-Network Model is algebraic rather than bibliometric: neighboring actors need not have identical private beliefs so long as their expressed discourse is compatible after edge-specific transformations (Hansen et al., 2020). This formulation directly represents selective opinion modulation and lying (Hansen et al., 2020).

In transformer interpretability, discourse is formalized as a sparse causal computation graph. “Discursive circuits” are defined operationally by edge-level causal interventions, with importance score

$g(e) = L(x_{cf} \mid do(E=e_{ori})) - L(x_{cf}),$

and an attribution approximation

$$g(e) \approx (z_u^{ori} - z_u^{cf})^\top \nabla_{v} L(x_{cf}),$$

for edge $e=(u,v)$ (Miao et al., 13 Oct 2025). This is a mechanistic neural version of the Discursive-Network Model in which discourse relations are processed by a sparse subgraph of attention heads and MLP blocks rather than diffusely throughout the model (Miao et al., 13 Oct 2025).

A more abstract network epistemology appears in work on human–LLM systems, where a discursive network is formally defined as

$H_x = - \sum_x p_x \log_2 p_x$0

with actors, statements, belief sets, goal functions, communications, invalidations, persuasion functions, and update rules (Gutiérrez, 9 Jul 2025). In its reduced two-state form, truth and falsehood evolve according to hazards of drift, self-repair, fabrication, and external detection (Gutiérrez, 9 Jul 2025). Although this is not a semantic network in the classical sense, it treats reliability as an emergent network property of discursive interaction.

5. Methodological implementations

The methodological repertoire of discursive-network modeling is heterogeneous but structurally consistent. In the original scientific-literature program, units of analysis can include documents, cited references, title words, author names, institutional addresses, journals, and combinations of these (0911.1308). Relevant relations include citation links, co-citation, co-word co-occurrence, journal-journal citations, clustering among publications, and trajectories over time (0911.1308). Matrices can be constructed as documents-by-words, documents-by-references, journals-by-journals, or three-way arrays across dimensions and time (0911.1308).

Leydesdorff’s visualization program generalizes this into a matrix-based workflow in which documents serve as common units for combining variables such as words, authors, references, institutions, or countries (Leydesdorff, 2012). The framework explicitly recommends asymmetrical matrices such as document-author and document-word matrices and their concatenation into multimodal structures (Leydesdorff, 2012). The latent structure is then reconstructed through factor analysis, SVD, cosine similarity, clustering, modularity, and dynamic layout optimization (Leydesdorff, 2012).

In social-media research, one major methodological branch begins with bipartite incidence and then validates projections by a maximum-entropy null model. In the study of X/Twitter discursive communities, a bipartite authors–audience retweet network is projected onto the verified-author layer, and only statistically significant shared-audience overlaps are retained under the Bipartite Configuration Model (Guarino et al., 2024). Community detection is then applied to that validated projection, and labels are propagated back to the broader user base (Guarino et al., 2024). The resulting communities are interpreted as audience-alignment clusters around elite communicators.

Another branch begins from semantic clustering and then induces actor proximity from shared discursive participation. In cross-platform narrative prediction, each user $H_x = - \sum_x p_x \log_2 p_x$1 is represented by a TF-IDF-weighted vector over narratives: $H_x = - \sum_x p_x \log_2 p_x$2 and user-user similarity is computed by cosine similarity

$H_x = - \sum_x p_x \log_2 p_x$3

This produces a platform-invariant discourse network in which proximity predicts cross-platform narrative emergence (Gerard et al., 10 Oct 2025). A closely related design is used to reconstruct cross-platform discourse networks in fragmented ecosystems and identify “bridge users” from shared participation in latent narratives (Gerard et al., 22 May 2025).

A third branch is triadic and context-sensitive rather than projection-based. In the Joint Attention-Interaction-Creation framework, collaborative discourse is modeled by three linked networks: an Attention Network weighted by semantic similarity of jointly referenced quotes, an Interaction Network weighted by frequency of direct conversational ties, and a Creation Network derived from a projected student-word incidence graph (Zhu et al., 2023). This extends the Discursive-Network Model beyond actor–concept incidence toward a layered representation of discourse origins, interpersonal exchange, and semantic production.

6. Applications and domain-specific variants

The Discursive-Network Model has been applied most directly to scientific communication, where it is used to study codification, diffusion trajectories, specialty formation, journal structures, co-word relations, and socio-cognitive regimes (0911.1308). Because publications, references, and textual attributes are traceable, the model treats the self-organization of scientific knowledge as measurable (0911.1308).

In international relations, unsupervised topic modeling combined with speaker-topic network analysis has been used to reconstruct the “discursive landscape” of UN Security Council debate, showing that actor-topic incidence can recover both issue dynamics and coalition structure (Schoenfeld et al., 2018). In this setting, the model is less about codification than about thematic alignment, issue ownership, and the relation between discourse and geopolitical blocs (Schoenfeld et al., 2018).

In online political communication, retweet-based discursive communities and hashtag-based semantic networks have been combined to study the uneven exposure of communities to d/misinformation narratives (Mattei et al., 2021). Related work shows that many Twitter discursive communities exhibit a bow-tie topology composed of IN, SCC, OUT, TUBES, INTENDRILS, OUTTENDRILS, and OTHERS, with the SCC and SCC-to-OUT flows especially implicated when low-quality content is present (Mattei et al., 2022). This yields a mesoscale model of public debate in which producers, mutually reinforcing cores, and audiences are topologically differentiated (Mattei et al., 2022).

In Web3 ecosystems, transaction graphs and off-chain discourse are integrated into a sociotechnical account in which on-chain behavior creates relational structure, relational structure shapes visibility and influence, and off-chain narrative production either consolidates durable communities or leaves activity fragmented and transactional (Kuskova et al., 20 Apr 2026). This suggests that discursive-network models can link economic exchange, discourse, identity, and inequality within the same analytical frame (Kuskova et al., 20 Apr 2026).

In generative AI research, the model has bifurcated into at least two directions. One treats large generative systems as Large Discourse Models operating on “the discursive projection” of human experience sedimented in documents, with an ontological triad of phenomenal world, embodied cognition, and documentary-discursive sedimentation (Lakel, 22 Dec 2025). The other treats mixed human–LLM systems as discursive networks in which people and models are equal nodes exchanging, invalidating, and revising statements (Gutiérrez, 9 Jul 2025). These are conceptually distinct, but both extend the model from discourse about science to discourse produced by or through synthetic agents.

7. Controversies, limitations, and conceptual scope

A recurring misconception is that the Discursive-Network Model is merely a co-occurrence graph of words. The literature consistently rejects that reduction. In the scientific-communication tradition, discourse includes references, citations, journal structures, and specialty-specific codes as well as lexical patterns (0911.1308). In the vector-space tradition, latent semantic organization is analytically distinct from observable ties (Leydesdorff, 2012). In social-media variants, actor proximity may be induced from shared narrative participation rather than direct interaction (Gerard et al., 22 May 2025, Gerard et al., 10 Oct 2025).

Another misconception is that discursive-network approaches are equivalent to classical diffusion models. Several papers explicitly argue otherwise. Lucio-Arias and Leydesdorff distinguish bottom-up diffusion from top-down codification and insist that later discourse reconstructs earlier texts “against the axis of time” (0911.1308). Cross-platform narrative prediction likewise frames emergence as a network proximity problem rather than direct cross-platform contagion (Gerard et al., 10 Oct 2025). A plausible implication is that many discursive-network models are better understood as models of structurally conditioned uptake than of literal transmission.

The family resemblance of these models should also not obscure their differences. Some are strongly interpretive and synthetic, as in the systems-theoretical model of scientific discourse (0911.1308). Some are algebraic and dynamical, as in discourse sheaves (Hansen et al., 2020). Some are probabilistic latent-variable models, as in coupled distributed topics (Srivatsan et al., 2018). Some are primarily empirical pipelines built from clustering, TF-IDF weighting, and nearest-neighbor search (Gerard et al., 22 May 2025, Gerard et al., 10 Oct 2025). It would therefore be misleading to speak of a single standardized formalism.

A major limitation in several empirical variants is that network proximity or co-participation does not directly encode stance. Topic-based or narrative-based incidence can show that two actors inhabit the same issue space without showing whether they endorse or oppose the same claims (Schoenfeld et al., 2018, Gerard et al., 22 May 2025). Likewise, hashtag co-usage can indicate semantic alignment or common attention rather than agreement (Mattei et al., 2021). This suggests that actor–concept models without polarity capture shared orientation, not necessarily shared position.

There are also domain-specific constraints. In some AI-oriented work, the model remains more programmatic than formally explicit about graph semantics or edge types (Lakel, 22 Dec 2025). In misinformation-as-gossip theory, the account is deliberately conceptual and does not provide equations (Bourbon et al., 2022). In mechanistic interpretability, the circuits framework is limited to transformer architectures and controlled discourse tasks (Miao et al., 13 Oct 2025). These limits do not negate the model family, but they delimit what each instance can claim.

8. Synthesis

Across its major formulations, the Discursive-Network Model treats discourse as a self-organizing, relationally structured, and often recursively updated domain in which units of communication acquire meaning through their positions in networks and through latent semantic organization. In scientific communication, this means that publications, references, words, journals, and specialties form a second-order textual dynamics whose codification can be measured (0911.1308). In vector-space approaches, it means that observable ties and latent dimensions must be analyzed jointly rather than conflated (Leydesdorff, 2012). In computational social-science variants, it means that actor proximity can be reconstructed from shared participation in topics, hashtags, narratives, or reply-linked latent states (Schoenfeld et al., 2018, Mattei et al., 2021, Srivatsan et al., 2018, Gerard et al., 22 May 2025).

The model’s enduring contribution is the relocation of explanation from isolated agents or isolated texts to structured relations among discursive traces. Whether instantiated through entropy decomposition, sheaf Laplacians, bipartite projections, sparse computational graphs, or narrative-affiliation networks, the same principle recurs: discourse is not an epiphenomenon laid over a preexisting network, but a networked process in its own right, with its own retention structures, transformation rules, control mechanisms, and emergent regimes (0911.1308, Hansen et al., 2020, Miao et al., 13 Oct 2025). This suggests that the most general definition of the Discursive-Network Model is a model in which discourse is simultaneously the medium, the object, and the architecture of social or intellectual organization.